Enhancing Earth system models efficiency: Leveraging the Automatic Performance Profiling tool
Roc Salvador Andreazini, Xavier Yepes-Arbós, Stella Valentina Paronuzzi Ticco, Oriol Tintó Prims, Mario Acosta · 2025
Earth system models (ESMs) are essential to understand and predict climate variability and change. However, their complexity and computational demands of high-resolution simulations often lead to performance bottlenecks that can impede research progress. Identifying and resolving these inefficiencies typically require significant expertise and manual effort, posing challenges for both climate scientists and High Performance Computing (HPC) engineers.We propose automating performance profiling as a solution to help researchers concentrate on improving and optimizing their models without the complexities of manual profiling. The Automatic Performance Profiling (APP) tool brings this solution to life by streamlining the generation of detailed performance reports for climate models.The tool ranges from high-level performance metrics, such as Simulated Years Per Day (SYPD), to low-level metrics, such as PAPI counters and MPI communication statistics. This dual-level reporting makes the tool accessible to a wide range of users, from climate scientists seeking a general understanding of the model efficiency, to HPC experts requiring granular insights for advanced optimizations.Seamlessly integrated with Autosubmit, the workflow manager developed at the Barcelona Supercomputing Center (BSC), APP ensures compatibility with complex climate modelling workflows. By automating the collection and reporting of key metrics, APP reduces the effort and expertise needed for performance profiling, empowering users to enhance the scalability and efficiency of their climate models.APP currently supports multiple models, including the EC-Earth4 climate model and the NEMO ocean model, and is compatible with different HPC systems, such as Marenostrum 5 and ECMWF’s supercomputer. Furthermore, the modular design of the tool allows adding new models and HPC platforms easily.